Are More Steps Better? Differences Between Karras and Beta Schedulers

Introduction

Hello, this is Easygoing.

Today, let’s explore the concepts of steps and schedulers in image generation AI.

Are More Steps Better?

Image generation involves gradually removing noise from a noisy image to create an illustration.

The number of steps determines how many times noise is removed. Generally, more steps are said to produce higher-quality images.

So, what are the advantages of increasing the number of steps?

Think of Steps as Division!

Steps should be thought of in terms of division, not addition.

Let’s look at an example of generating an illustration with 5 steps.

5 Steps

In ComfyUI, setting 5 steps automatically adds 1 extra step, resulting in a total of 6 steps.

With fewer steps, the gaps between steps are larger, leading to greater errors between steps.

Larger errors result in less accurate rendering, producing illustrations that are blurry with fewer details.

Additionally, insufficient noise removal in the final step leaves residual noise in the completed illustration.

10 Steps

Now, let’s try generating an illustration with 10 steps.

Increasing the number of steps reduces errors between steps, improving calculation accuracy and adding more details to the illustration.

It also reduces residual noise, resulting in a clear and vibrant illustration.

Are More Steps Always Better?

Let’s try increasing the step count even further.

20 Steps

30 Steps

50 Steps

Excessively increasing the step count can reduce details and make colors appear dull or burnt.

Depending on the model, too many steps can actually be detrimental.

What’s the Optimal Step Count?

The optimal step count varies by model.

Local image generation is primarily divided into three architectures:

gantt
title e-pred and v-pred
dateFormat YYYY-MM-DD
tickInterval 6month
section e-pred
Stable Diffusion 1 :done, a1, 2022-08-22, 2025-06-15
section v-pred
Stable Diffusion 2.0 : b1, 2022-11-24, 2025-06-15
section e-pred + v-pred
Stable Diffusion XL 1.0 :done, c2, 2023-07-27, 2025-06-15
section Flow-matching         
Stable Diffusion 3 : d1, 2024-06-12, 2025-06-15
AuraFlow : d2, 2024-07-12, 2025-06-15
Flux.1   : d3, 2024-08-01, 2025-06-15
HiDream-I1   : d4, 2025-04-06, 2025-06-15

• e-pred: SD 1.5 / SDXL
• v-pred: Some SDXL models
• Flow-matching: SD 3.5 / AuraFlow / Flux.1 / HiDream

e-pred: SD 1.5 / SDXL

e-pred is the conventional method used in Stable Diffusion 1.5 and Stable Diffusion XL.

v-pred: Some SDXL Models

v-pred is used in some anime models of Stable Diffusion XL, enabling efficient noise removal and image generation with fewer steps.

Flow-matching: SD 3.5 / AuraFlow / Flux.1 / HiDream

Flow-matching is used in next-generation image generation models starting with Stable Diffusion 3. Like v-pred, it removes noise efficiently but involves broader architectural changes.

SD 1.5 / SDXL: 20–40 Steps

e-pred requires sufficient steps for standard noise removal.

The official Stability AI Hugging Face SDXL_Base Diffusers pipeline sets the default step count to 40 steps, which serves as a benchmark for SDXL.

However, recent advancements in samplers and schedulers have improved noise removal accuracy, allowing 20–30 steps to achieve sufficient quality.

Custom models with additional training may specify recommended step counts, which should generally be followed.

• Animagine-XL 4.0 Opt recommends 25–28 steps.

Next-Generation Models: 10–20 Steps

v-pred, used in some SDXL anime models, removes noise more efficiently than e-pred.

Flow-matching, adopted by next-generation AI models, removes noise with even higher precision.

Both v-pred and Flow-matching often achieve satisfactory quality with 10–20 steps.

Models that remove noise aggressively may experience color burn with excessively high step counts. It’s best to aim for the fewest steps possible while maintaining quality.

Understanding Schedulers!

Let’s dive into schedulers in the second half.

Schedulers can feel complex, but we’ll break them down with visuals.

Early Stages for Composition, Final Stages for Details

When generating an illustration, the initial composition is established first, followed by increased detailing in the middle, and finalized with fine details.

The initial phase significantly impacts quality, so minimizing errors in this stage is crucial.

Simple Scheduler

Let’s start with the conventional simple scheduler.

The simple scheduler removes noise aggressively in the initial phase, which can lead to errors in the critical early stage.

Karras Scheduler

The karras scheduler reduces initial noise removal to minimize errors in the early phase.

Instead, it removes more noise in the middle phase, adjusting to leave less residual noise than the simple scheduler.

The karras scheduler can generate stable, high-quality illustrations with fewer steps.

Beta Scheduler

The beta scheduler is designed to use all steps efficiently, resulting in a smooth noise reduction curve.

It retains more noise in the middle phase, which enhances illustration details.

However, without a model that removes noise efficiently, residual noise may remain in the final illustration.

Measurement Data

Here are the noise reduction tables and graphs measured in ComfyUI for each scheduler.

So, What’s Recommended?

• For fast generation with SD1.5/SDXL ➡ karras
• For enhancing details with next-generation models ➡ beta

Conclusion: Try the Beta Scheduler!

• SDXL: 20–40 steps
• Next-generation models: 10–20 steps
• Recommended: beta scheduler

The optimal step count and scheduler vary by model.

Recent high-parameter models like HiDream or Flux.1 can produce stunning details with the beta scheduler.

Lately, I believe generating images with high-precision models and samplers using as few steps as possible is the way to go.

Thank you for reading until the end!